Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
Control Theory and Technology http://link.springer.com/journal/11768
Infinite horizon indefinite stochastic linear quadratic control for discrete-time systems Weihai ZHANG 1† , Yan LI 1 , Xikui LIU 2 1.College of Electrical Engineering and Automation, Shandong University of Science and Technology, Qingdao Shandong 266590, China; 2.College of Mathematics and Systems Science, Shandong University of Science and Technology, Qingdao Shandong 266590, China Received 18 October 2014; revised 10 July 2015; accepted 13 July 2015
Abstract This paper discusses discrete-time stochastic linear quadratic (LQ) problem in the infinite horizon with state and control dependent noise, where the weighting matrices in the cost function are assumed to be indefinite. The problem gives rise to a generalized algebraic Riccati equation (GARE) that involves equality and inequality constraints. The well-posedness of the indefinite LQ problem is shown to be equivalent to the feasibility of a linear matrix inequality (LMI). Moreover, the existence of a stabilizing solution to the GARE is equivalent to the attainability of the LQ problem. All the optimal controls are obtained in terms of the solution to the GARE. Finally, we give an LMI -based approach to solve the GARE via a semidefinite programming. Keywords: Indefinite stochastic LQ control, discrete-time stochastic systems, generalized algebraic Riccati equation, linear matrix inequality, semidefinite programming DOI 10.1007/s11768-015-4147-x
1 Introduction It is well known that the linear quadratic (LQ) optimal control problem of deterministic systems was first founded by Kalman [1], which has been playing important role in both theory and applications. The deterministic LQ problem has been discussed extensively by many researchers; see, e.g., [2–5]. The stochastic LQ
problem was initiated by Wonham [6] and has been investigated; see [7–14] and the references therein. In the literature, it is a common assumption that the control weighting matrix should be positive definite and the state weighting matrix should be nonnegative. In this case, the solvability of the LQ problem is equivalent to that of the Riccati equations. However, a class of stochastic LQ problems with indefinite control weights
† Corresponding author. E-mail:
[email protected]. This work was supported by the National Natural Science Foundation of China (Nos. 61174078, 61170054, 61402265) and the Research Fund for the Taishan Scholar Project of Shandong Province of China.
© 2015 South China University of Technology, Academy of Mathematics and Systems Science, CAS, and Springer-Verlag Berlin Heidelberg
231
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
may still be well-posed [15]. The solvability of indefinite stochastic LQ problem is closely linked to the solvability of the indefinite stochastic Riccati equations. There are many works focusing on this issue, we refer the reader to [16–18]. For the discrete-time LQ control, Y. Huang et al. [19, 20] studied a class of special cases, where the system is described by a difference equation with control and state dependent noise. In [21], the optimal control is obtained for the systems with only control dependent noise. These papers dealt with the LQ problem with the positive definite control weighting matrices in the cost functional. A discrete-time indefinite LQ control in a finite horizon with state and control dependent noise is studied in [22]. Analytical properties of the constrained discrete-time indefinite stochastic LQ control in finite time horizon were extensively studied in [23–25]. This paper considers the discrete-time indefinite stochastic LQ problem in the infinite horizon. Different from the finite horizon case, in order to guarantee the well-posedness of the LQ problem and the existence of the feedback stabilizing control, we have to define some concepts such as stabilizability. A generalized algebraic Riccati equation (GARE) is introduced. It turns out that the attainability of the LQ problem is necessary and sufficient for the existence of the stabilizing solution to the GARE. Meanwhile, we introduce a linear matrix inequality (LMI) condition and show that the well-posedness of the LQ problem is equivalent to the feasibility of the LMI. Furthermore, we present all optimal controls via the solution to the GARE. Finally, we give an LMI-based approach to solve the GARE via a semidefinite programming. The remainder of this paper is organized as follows. In Section 2, we present the notions of stabilizability and some preliminaries. Section 3 shows that the solvability of the GARE is sufficient for the well-posedness of the LQ problem and the existence of an optimal control. Section 4 contains main results of the paper. In Section 5, we give an LMI-based approach to solve the GARE via a semidefinite programming. Section 6 ends this paper with some concluding remarks. For convenience, we adopt the following notations. M represents the transpose of a matrix M; Tr(M) is the trace of a square matrix M; M > 0 (M 0) means that M is a positive definite (positive semi-definite) symmetric matrix; E[x] represents the mathematical expectation of a random variable x; Rk is the k-dimensional Euclidean space with the usual 2-norm · ; Rm×n is the vector
space of all m × n matrices with entries in R; M† means the Moore-Penrose pseudo inverse of a matrix M; I is the identity matrix with an appropriate dimension; Sn denotes the set of all real n × n symmetric matrices; and N = {0, 1, 2, . . .}.
2 Problem formulation and preliminaries Consider the following stochastic discrete-time system of the form: xk+1 = Axk +Buk +[Cxk +Duk ]wk , k ∈ N, x0 ∈ Rn , (1) where A, B, C and D are constant matrices with appropriate dimensions. x ∈ Rn is called the system state, u ∈ Rm is the control input. x0 ∈ Rn is the initial state which is deterministic. {wk }k0 are the one-dimensional independent random variables defined on the complete probability space (Ω, F , P), such that E[wk ] = 0 and E[ws wt ] = δst , where δst is the Kronecker delta. We denote Fk the σ-algebra generated by wk , k ∈ N, i.e., Fk = σ( ws : 1 s k). Let L2 (Ω, Rm ) represent the space of Rm -valued random vectors ξ with Eξ2 < ∞. l2w (N, Rm ) consists of all sequences y = {yk : yk ∈ Rm }k∈N , such that yk ∈ L2 (Ω, Rm ) is Fk−1 measurable for k ∈ N, where we define F−1 = {φ, Ω}, i.e., y0 is a constant. The l2 -norm of y ∈ l2w (N, Rm ) is ∞ defined by y2l2 (N,Rm ) = E[yk 2 ]. w
k=0
For simplicity of our discussion, we give the following definitions. Definition 1 Consider system (1) with uk = 0. System (1) is said to be mean-square stable [19] if for any x0 ∈ Rn , the corresponding state satisfies lim Exk 2 = k→∞
0. u = {uk : k ∈ N} (may be an open-loop control) is said to be a mean-square stabilizing control (with respect to x0 ) if the corresponding state xk of (1) satisfies lim Exk 2 = 0. u = {uk : k ∈ N} with uk = Kxk is called k→∞
a mean-square feedback stabilizing control law if for every x0 , the closed-loop system xk+1 = (A + BK)xk + (C + DK)xk wk , k ∈ N, x0 ∈ Rn is mean-square stable, where K is a constant matrix. Definition 2 System (1) is said to be stabilizable in the mean square sense if there exists a mean-square feedback stabilizing control law uk = Kxk , where K is a constant matrix.
232
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
For system (1), the admissible control set Uad (x0 ) is defined as follows:
Lemma 2 [26] Let a symmetric matrix M be given. Then
Uad (x0 ) ={u ∈ l2w (N, Rm )|uk is mean-square
M† = M† , MM† = M† M,
stabilizing with respect to a given x0 }.
M 0 if and only if M† 0.
For any (x0 , uk ) ∈ Rn × Uad (x0 ), the associated cost functional to system (1) is defined as J(x0 , u) =
∞ k=0
E[xk Qxk + 2xk Luk + uk Ruk ],
(2)
where L ∈ Rn×m , Q ∈ Sn and R ∈ Sm are given matrices. The LQ optimal control problem is to find a control sequence u∗ = (u∗0 , . . . , u∗n , . . .) ∈ Uad (x0 ) such that J(x0 , u∗ ) = V(x0 ) =
inf
u∈Uad (x0 )
J(x0 , u).
(3)
We call V(x0 ) the optimal cost value. Definition 3 The LQ problem is called well-posed if −∞ < V(x0 ) < +∞, ∀x0 ∈ Rn . A well-posed LQ problem is called attainable if there exists a control sequence (u∗0 , . . . , u∗n . . .) that achieves V(x0 ). We suppose that system (1) is stabilizable throughout this paper. Hence, Uad (x0 ) is nonempty for each x0 . Now, we present a new GARE as follows. Definition 4 The constrained algebraic equation on P ∈ Sn ⎧ ⎪ † ⎪ ⎨ F(P) − H (P)G (P)H(P) = 0, (4) ⎪ ⎪ ⎩ G(P)G† (P)H(P) − H(P) = 0, G(P) 0 with
⎧ ⎪ ⎪ F(P) = A PA + C PC − P + Q, ⎪ ⎪ ⎪ ⎨ H(P) = B PA + D PC + L , ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ G(P) = R + B PB + D PD
(5)
is called a constrained GARE.
†
Lemma 4 (Extended Schur’s lemma) [27] Let matrices M = M , N, R = R be given with appropriate sizes. Then the following conditions are equivalent: i) M − NR† N 0, R 0, and N(I − RR† ) = 0. ⎤ ⎡ ⎢⎢⎢ M N ⎥⎥⎥ ii) ⎢⎢⎣ ⎥⎥⎦ 0. N R ⎤ ⎡ ⎢⎢ R N ⎥⎥ ⎥⎥ 0. ⎢ iii) ⎢⎢⎣ ⎥ N M⎦
3
Sufficiency of the GARE
In this section, it is shown that the solvability of the GARE (4) is sufficient for the well-posedness of the LQ problem and the existence of an optimal control. Moreover, we show that any optimal control can be determined by means of the solution to the GARE. Theorem 1 If the GARE (4) admits a solution P and there exist Yk ∈ Rm×n and Zk ∈ Rm such that the following control: (Yk , Zk )
uk
= −(G† (P)H(P) − Yk + G† (P)G(P)Yk )xk −G† (P)G(P)Zk + Zk
(6)
is admissible for any initial x0 . Then LQ problem (1)–(3) (Y , Z ) is attainable. Furthermore, uk k k is the optimal control and the optimal cost value is uniquely determined by V(x0 ) = x0 Px0 .
Let us give some lemmas needed in the proof of our main results. Lemma 1 [26] For any matrix M ∈ Rm×n , there is a unique matrix M† ∈ Rn×m , which satisfies †
Lemma 3 [22] Let matrices L, M, N be given, then the matrix equation LXM = N has a solution X if and only if LL† NMM† = N. X is given by X = L† NM† + Y − L† LYMM† , where Y is a matrix with an appropriate dimension.
†
†
MM M = M , M MM = M , (MM† ) = MM† , (M† M) = M† M. M† is called the Moore-Penrose pseudo inverse of M.
Proof Let P solve the GARE (4). It is clear that ∀T ∈ N, P ∈ Sn , E[xT PxT − x0 Px0 ] =
T−1
E(xk+1 Pxk+1 − xk Pxk ) ⎡ ⎤ ⎡ ⎤ T−1 ⎢⎢xk ⎥⎥ ⎢⎢⎢xk ⎥⎥⎥ E ⎢⎢⎣ ⎥⎥⎦ Q(P) ⎢⎢⎢⎣ ⎥⎥⎥⎦ , = uk uk k=0 k=0
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
is admissible, where xk is the solution to (1) under the above control with the initial state x0 .
where ⎤ ⎡ ⎢⎢A PA + C PC − P A PB + C PD⎥⎥ ⎥⎥ . ⎢ Q(P) = ⎢⎢⎣ ⎥ B PA + D PC B PB + D PD⎦
The following definition is concerned with the maximal solution. Definition 6 A matrix P is called a maximal solution to the GARE (4) if P P∗ for any P∗ satisfying ⎧ ⎪ ∗ ∗ † ∗ ∗ ⎪ ⎨ F(P ) − H (P )G (P )H(P ) 0, (10) ⎪ ⎪ ⎩ H(P∗ ) − G(P∗ )G† (P∗ )H(P∗ ) = 0, G(P∗ ) 0.
Let T → ∞, then ⎡ ⎤ ⎡ ⎤ ⎢⎢xk ⎥⎥ ⎢⎢xk ⎥⎥ ⎥ ⎢ E ⎢⎢⎣ ⎥⎥⎦ Q(P) ⎢⎢⎢⎣ ⎥⎥⎥⎦ = −x0 Px0 . uk uk k=0 ∞
By adding the above equality to the performance index, we have J(x0 , u) =
∞ k=0
E[xk (Q + A PA + C PC − P)xk
+ 2xk (A PB + C PD + L)uk
+ uk (R + B PB + D PD)uk ] + x0 Px0 ∞ = E[xk F(P)xk + 2xk H (P)uk k=0
+ uk G(P)uk ] + x0 Px0 .
(7)
Define M1k
†
= G (P)G(P)Yk − Yk ,
M2k
= G (P)G(P)Zk − Zk .
By completing squares, (7) can be rewritten as
k=0
E[(uk + (G† (P)H(P) + M1k )xk
+ M2k ) G(P)(uk + (G† (P)H(P) + M1k )xk + M2k )] + x0 Px0 .
(8)
This implies that the control sequence (Yk , Zk )
uk = uk
= −[(G† (P)H(P) + M1k )xk + M2k ], k ∈ N
minimizes J with the optimal value given by
x0 Px0 .
Proof Assume that P1 and P2 are different stabilizing solutions to (4). By Theorem 1, it follows that x0 P1 x0 = x0 P2 x0 for any x0 , so P1 = P2 . Let any P∗ satisfy (10) and P be the stabilizing solution to (4). Putting P∗ in (8) we assert that
By Theorem 1, it is easy to show that x0 Px0 = V(x0 ) x0 P∗ x0 . Therefore, P is a maximal stabilizing solution to the GARE due to Definition 6.
G(P)M1k = 0, G(P)M2k = 0.
∞
By Definition 6, it is clear that the maximal solution must be unique if it exists. Now, let us turn to the GARE (4). Theorem 2 There is at most one stabilizing solution to (4). Moreover, a stabilizing solution to (4) is also its maximal solution.
J(x0 , u) x0 P∗ x0 , ∀u ∈ Uad (x0 ).
†
Hence, we can obtain that
J(x0 , u) =
233
Definition 5 A solution P to the GARE (4) is called stabilizing if there exists an admissible control determined by (6).
The following corollaries are special cases of the above result. Corollary 1 Suppose that the GARE (4) admits a stabilizing solution P. If G(P) = 0, then any admissible control is optimal and the GARE (4) reduces to ⎧ ⎪ ⎪ ⎪ ⎪ P = A PA + Q + C PC, ⎪ ⎨ (11) B PA + D PC + L = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ R + B PB + D PD = 0. Proof By (7) and G(P) = 0, we can show that J(x0 , u) = x0 Px0 , which implies that V(x0 ) = x0 Px0 for any uk ∈ Uad (x0 ).
Remark 1 A solution P to the GARE (4) is stabilizing if and only if for any x0 there exists some Zk ∈ Rm such that the following control:
Corollary 2 Let P be a stabilizing solution to the GARE (4). If G(P) > 0, then the LQ problem(1)–(3) is uniquely solvable. The unique optimal control is given by
uk = −G† (P)H(P)xk + [I − G† (P)G(P)]Zk
uk = −G−1 (P)H(P)xk .
(9)
234
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
Proof Using Theorem 1, we immediately obtain Corollary 2.
4 Well-posedness and attainability of LQ problem In this section, we first present the connection between the well-posedness of the LQ problem and the solvability of the GARE. Then, we study the wellposedness via the LMI condition. Finally, we establish the link between the attainability of the LQ problem and the solvability of the GARE. Lemma 5 The LQ problem (1)–(3) is well-posed if and only if there exists a symmetric constant matrix P such that V(x0 ) = x0 Px0 , ∀x0 ∈ Rn .
(12)
Proof (12) can be shown by a simple adaptation of the well-known result in the deterministic case [2]. We introduce the following convex set P of Sn : ⎤ ⎡
⎢ F(P) H (P)⎥⎥ n ⎢ ⎥⎥ 0 . ⎢ (13) P = P ∈ S | ⎢⎢⎣ ⎥⎦ H(P) G(P) Theorem 3 The LQ problem (1)–(3) is well-posed if and only if the set P is nonempty. Moreover, there ˜ ∀P˜ ∈ P. exists P ∈ P such that P P, Proof (Sufficiency) Assume that the set P is nonempty, let ∀P˜ ∈ P. Then, adding the following equality: +∞ k=0
˜ k+1 − x Px ˜ k ] = −E[x Px ˜ 0] E[xk+1 Px 0 k
+∞ k=0
x0 Px0
h k=0
E[xk Qxk + 2xk Luk + uk Ruk
+ xh+1 Pxh+1 ], ∀h 0, ∀u ∈ Uad (x0 ). Based on (1) and the above inequality, we conclude that E[xh+1 Pxh+1 ] − x0 Px0 + =
h
k=0 h k=0
E[xk Qxk + 2xk Luk + uk Ruk ] E[xk+1 Pxk+1 − xk Pxk
+ xk Qxk + 2xk Luk + uk Ruk ] ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎢⎢xk ⎥⎥ ⎢⎢ F(P) H (P)⎥⎥ ⎢⎢xk ⎥⎥ h ⎢ ⎥⎥ ⎢⎢ ⎥⎥ 0. ⎥ ⎢ E ⎢⎢⎣ ⎥⎥⎦ ⎢⎢⎣ = ⎥⎢ ⎥ uk H(P) G(P) ⎦ ⎣uk ⎦ k=0
(15)
Setting uk = u¯ and letting h = 0, we have ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎢⎢x0 ⎥⎥ ⎢⎢ F(P) H (P)⎥⎥ ⎢⎢x0 ⎥⎥ ⎥⎥ ⎢⎢ ⎥⎥ ⎢⎢⎢ ⎥⎥⎥ ⎢⎢⎢ ⎣ u¯ ⎦ ⎣H(P) G(P) ⎥⎦ ⎢⎣ u¯ ⎥⎦ 0. Because x0 and u¯ are arbitrary, it is easy to see that ⎤ ⎡ ⎢⎢ F(P) H (P)⎥⎥ ⎥⎥ ⎢⎢⎢ ⎣H(P) G(P) ⎥⎦ 0. This means that P ∈ P. Employing (14), it follows that ˜ ∀P˜ ∈ P. P P,
to the cost function J(x0 , u) =
implies the well-posedness of the LQ problem. (Necessity) Assume that the LQ problem (1)–(3) is well-posed, Lemma 5 yields that there exists a symmetric matrix P such that V(x0 ) = x0 Px0 , ∀x0 ∈ Rn . By the dynamic programming principle, we obtain
E[xk Qxk + 2xk Luk + uk Ruk ]
and applying (1), we can see that for any (x0 , uk ) ∈ Rn × Uad (x0 ), ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎢⎢xk ⎥⎥ ⎢⎢ F(P) ˜ ⎥⎥ ⎢⎢xk ⎥⎥ ˜ H (P) ⎢ ⎥⎥ ⎢⎢ ⎥⎥ + x Px ⎥ ⎢ ˜ 0 J(x0 , u) = E ⎢⎢⎣ ⎥⎥⎦ ⎢⎢⎣ 0 ˜ G(P) ˜ ⎥⎦ ⎢⎣u ⎥⎦ u H(P) k=0 +∞
k
k
˜ 0. x0 Px Since x0 and uk are arbitrary, thus ˜ 0 V(x0 ) x0 Px
(14)
The following theorem can be viewed as the converse of Theorem 1, which plays an essential role in this paper. Theorem 4 The LQ problem (1)–(3) is attainable for any x0 , then the GARE (4) has a stabilizing solution. Moreover, any optimal control is given by (6). Proof If the LQ problem (1)–(3) is attainable, it is also well-posed. By Theorem 5, there exists a maximal element P ∈ P satisfying V(x0 ) = x0 Px0 and ⎤ ⎡ ⎢⎢ F(P) H (P)⎥⎥ ⎥⎥ 0. ⎢⎢ ⎥ ⎢⎣ H(P) G(P) ⎦
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
Using Lemma 4, it is obvious that ⎧ † ⎪ ⎨ F(P) − H (P)G (P)H(P) 0, ⎪ ⎩ H (P)(I − G(P)G† (P)) = 0, G(P) 0.
(16)
Let u∗ be an optimal control for any initial x0 . As same as Theorem 5, the following (17) is derived: V(x0 ) =J(x0 , u∗ ) +∞ = E[xk (F(P) − H (P)G† (P)H(P))xk ] k=0 +∞
+
×
k=0 (u∗k
min c x s.t. F(x) = F0 + (17)
Combining (16) with V(x0 ) = x0 Px0 , it can be shown that +∞ k=0 +∞ k=0
Together with (15), we obtain that P is a solution to the GARE (4). In what follows, we show that any optimal control u∗k can be given by (6). From (16), it yields that G(P)
1 2
[u∗k
†
+ G (P)H(P)xk ] = 0,
which implies G(P)u∗k
+
+ G(P)G (P)H(P)xk = 0.
By Lemma 3 with L = G(P), M = I, N = −G(P)G+ (P)H(P)xk , we solve the above equation and have the following solution u∗k = −G+ (P)H(P)xk + Y − G+ (P)G(P)Y. Thus u∗k can be represented by (6) with Zk = Y and Yk = 0. On the other hand, from Definition 5, it follows that P is a stabilizing solution to the GARE (4).
Characterizing LQ problem via SDP
In this section, we develop an approach based on semidefinite programming (SDP). We show that the sta-
i=1
xi Fi 0
(18)
Consider the following SDP problem: min −Tr(P) ⎤ ⎡ ⎢⎢ F(P) H (P)⎥⎥ ⎥⎥ 0. s.t. ⎢⎢⎢⎣ ⎥ H(P) G(P) ⎦
E[(u∗k +G† (P)H(P)xk ) G(P)(u∗k +G† (P)H(P)xk )] = 0.
F(P) − H (P)G† (P)H(P) = 0, k ∈ N.
m
is called a semidefinite programming (SDP). The SDP is feasible if there exists an x such that F(x) 0.
E[xk (F(P) − H (P)G† (P)H(P))xk ] = 0,
Hence, we have
5
bilizing of the feedback control can be examined via solving a SDP problem. We establish several relations among the GARE, the SDP and the optimality of the LQ problem. First, we introduce the following definition. Definition 7 [28] Let a vector c = (c1 , . . . , cm ) ∈ Rm and matrices F0 , F1 , . . . , Fm ∈ Sn be given. The following optimization problem:
E[(u∗k + G† (P)H(P)xk ) G(P) + G† (P)H(P)xk )] + x0 Px0 .
235
(19)
The following assertions provide connections among the well-posedness of the LQ problem, the feasibility of the SDP and the solvability of the GARE. Theorem 5 The SDP (19) is feasible if and only if the LQ problem (1)–(3) is well-posed. Proof sult.
By Theorem 5, we easily get the desired re
Theorem 6 If the SDP (19) is feasible, then it has a unique optimal solution P∗ satisfying V(x0 ) = x0 P∗ x0 , ∀x0 ∈ Rn . Proof By Theorem 5, it follows that the SDP (19) has a maximal solution P such that V(x0 ) = x0 Px0 , ∀x0 ∈ Rn , which is also an optimal solution to (19). Let P∗ be arbitrary optimal solution to (19). It is evident that Tr(P − P∗ ) = 0. Moreover, the maximality of P results in P − P∗ 0 . Therefore, P − P∗ = 0, i.e., P = P∗ . Theorem 7 If the LQ problem (1)–(3) is attainable, then the unique optimal solution to (19) is the stabilizing solution to the GARE (4). Proof From Theorems 2 and 6, the assertion is immediately obtained. Similar to the discussion in [29], we give a computational approach to determine the lower bound of the control weighting matrix R for the LQ problem to be well-posed. Definition 8 Let Q and L be given. The smallest
236
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015
r∗ ∈ R is called the well-posed margin if the LQ problem (1)–(3) is well-posed for any R r∗ I. Remark 2 By the above definition, if the smallest eigenvalue λmin (R) of R satisfies λmin (R) r∗ , then the LQ problem is well-posed. Otherwise, the LQ problem is ill-posed. In particular, if r∗ = 0, the LQ problem is ill-posed for any indefinite R. The following result shows that the well-posedness margin r∗ can be obtained numerically. Theorem 8 The well-posedness margin r∗ can be derived by solving the following SDP problem: min ⎡ r ⎤ ⎥⎥ ⎢⎢ F(P) (P) H ⎥⎥ 0. s.t. ⎢⎢⎢⎣ ⎥ H(P) B PB + D PD + rI⎦ Proof proved.
References [1] R. E. Kalman. Contributions to the theory of optimal control. Boletin de la Sociedad Matematica Mexicana, 1960, 5: 102 – 119. [2] B. D. O. Anderson, J. B. Moore. Optimal Control – Linear Quadratic Methods. New York: Prentice-Hall, 1989.
(20)
From Theorem 7, Theorem 8 is easily
Example 1 ing matrices:
with state and control dependent noise. A GARE has been introduced. The well-posedness of the LQ problem is equivalent to the feasibility of an LMI. Moreover, the attainability of the LQ problem is equivalent to the existence of a stabilizing solution to the GARE. All the optimal controls are obtained in terms of the solution to the GARE. To some extent, the results of this paper may be viewed as a discrete-time version of [29].
System (1)–(3) is specified by the follow-
[3] F. L. Lewis. Optimal Control. New York: John Wiley & Sons, 1986. [4] L. Ntogramatzidis, A. Ferrante. On the solution of the Riccati differential equation arising from the LQ optimal control problem. Systems & Control Letters, 2010, 59(2): 114 – 121. [5] A. Ferrante, L. Ntogrammatzidis. The generalized continuous algebraic Riccati equation and impulse-free continuous-time LQ optimal control. Automatica, 2014, 50(4): 1176 – 1180.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎢⎢−2 1⎥⎥ ⎢⎢ 1 ⎥⎥ ⎢⎢2 0⎥⎥ ⎢⎢ 1 ⎥⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎥⎥ , B = ⎢⎢ ⎥⎥ , C = ⎢⎢ ⎥⎥ , D = ⎢⎢⎢ ⎥⎥⎥ , A = ⎢⎢⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣0⎦ 0 1 0 0 0 ⎡ ⎤ ⎡ ⎤ ⎢⎢ 1 ⎥⎥ ⎢⎢−21 0 ⎥⎥ ⎥⎥ , L = 0, R = −1. x0 = ⎢⎢⎢⎣ ⎥⎥⎥⎦ , Q = ⎢⎢⎢⎣ ⎥ −1 0 6/5⎦
[6] W. M. Wonham. On a matrix Riccati equation of stochastic control. SIAM Journal on Control, 1968, 6: 312 – 326.
Solving the corresponding GARE (4) yields G = 5, ⎡ ⎤ ⎢⎢ 3 −2⎥⎥ ⎥⎥. Finally, we can calculate the H = [0 1], P = ⎢⎢⎢⎣ ⎥ −2 1 ⎦
[9] M. H. A. Davis. Linear Estimation and Stochastic Control. London: Chapman and Hall, 1977.
optimal control sequence uk = −[0 1/5]xk and the optimal cost value V(x0 ) = x0 Px0 = 8 . By Theorem 8, the well-posedness margin r∗ = 6. Remark 3 In stochastic systems, there is no similar definition to transfer function as in linear system theory. Therefore, we cannot define an optimal feedback control to be proper or improper as in [30]. In this paper, the admissible control set Uad (x0 ) is not limited to a static state feedback form, which may include other forms such as open-loop controls. However, from Theorem 4, if the LQ problem is attainable, all optimal controls take the form of (6), which completely characterizes all the optimal control forms.
[7] D. H. Jacobson, D. H. Martin, M. Pachter, et al. Extensions of Linear – Quadratic Control Theory. Lecture Notes in Control and Information Sciences. Berlin: Springer, 1980. [8] A. Bensoussan. Stochastic Control of Partially Observed Systems. Cambridge, U.K.: Cambridge University Press, 1992.
[10] E. Yaz. Infinite horizon quadratic optimal control of a class of nonlinear stochastic systems. IEEE Transactions on Automatic Control, 1989, 34(11): 1176 – 1180. [11] W. Zhang, B. Chen. On stabilizability and exact observability of stochastic systems with their applications. Automatica, 2004, 40(1): 87 – 94. [12] W. Zhang, H. Zhang, B. Chen. Generalized Lyapunov equation approach to state-dependent stochastic stabilization/detectability criterion. IEEE Transactions on Automatic Control, 2008, 53(7): 1630 – 1642. [13] Y. Hu, H. Jin, X. Zhou. Time-inconsistent stochastic linearquadratic control. SIAM Journal on Control and Optimization, 2012, 50(3): 1548 – 1572. [14] J. M. Yong. Linear-quadratic optimal control problems for meanfield stochastic differential equations. SIAM Journal on Control and Optimization, 2013, 51(4): 2809 – 2838.
6 Conclusions
[15] S. Chen, X. Li, X. Zhou. Stochastic linear quadratic regulators with indefinite control weight costs. SIAM Journal on Control and Optimization, 1998, 36(5): 1685 – 1702.
This paper has investigated the infinite horizon indefinite LQ control for discrete-time stochastic systems
[16] Z. Qian, X. Zhou. Existence of solutions to a class of indefinite stochastic Riccati equations. SIAM Journal on Control and Optimization, 2013, 51(1): 221 – 229.
W. Zhang et al. / Control Theory Tech, Vol. 13, No. 3, pp. 230–237, August 2015 [17] Q. Meng. Linear quadratic optimal stochastic control problem driven by a Brownian motion and a Poisson random martingale measure with random cofficients. Stochastic Analysis and Applications, 2014, 32(1): 88 – 109. [18] R. S. Burachik, C. Y. Kaya, S. N. Majeed. A duality approach for solving control-constrained linear-quadratic optimal control problems. SIAM Journal on Control and Optimization, 2014, 52(3): 1423 – 1456. [19] Y. Huang, W. Zhang, H. Zhang. Infinite horizon linear quadratic optimal control for discrete-time stochastic systems. Asian Journal of Control, 2008, 10(5): 608 – 615. [20] R. T. Ku, M. Athans. Further results on the uncertainty threshold principle. IEEE Transactions on Automatic Control, 1977, 22(5): 866 – 868. [21] A. Beghi, D. D. Alessandro. Discrete-time optimal control with control-dependent noise and generalized Riccati difference equations. Automatica, 1998, 34(8): 1031 – 1034. [22] M. Ait Rami, X. Chen, X. Zhou. Discrete-time indefinite LQ control with state and control dependent noises. Journal of Global Optimization, 2002, 23(3/4): 245 – 265. [23] G. Li, W. Zhang. Study on indefinite stochastic linear quadratic optimal control with inequality constraint. Journal of Applied Mathematics, 2013: DOI 10.1155/2013/805829. [24] W. Zhang, G. Li. Discrete-time indefinite stochastic linear quadratic optimal control with second moment constraints. Mathmatical Problems in Engineering, 2014: DOI 10.1155/2014/ 278142.
237
[28] L. Vandenberghe, S. Boyd. Semidefinite programming. SIAM Review, 1996, 38(1): 49 – 95. [29] M. Ait Rami, X. Zhou, J. B. Moore. Well-posedness and attainability of indefinite stochastic linear quadratic control in infinite time horizon. Systems & Control Letters, 2000, 41(2): 123 – 133. [30] B. R. Copeland, M. G. Safonov. Zero cancelling compensation for singular control problems and their application to the innerouter factorization problem. International Journal of Robust and Nonlinear Control, 1992, 2(2): 139–164. Weihai ZHANG received his Ph.D. degree from Zhejiang University in 1998, Hangzhou, China. He is currently a professor of Shandong University of Science and Technology. His research interests include stochastic optimal control, robust H∞ control, stochastic stability and stabilization. Email:
[email protected]. Yan LI received the M.Sc. and Ph.D. degrees from Shandong University of Science and Technology, China, in 2006 and 2015, respectively. She is a lecturer of Shandong University of Science and Technology. Her research interests include linear and nonlinear stochastic control. E-mail:
[email protected].
[25] X. Liu, Y. Li, W. Zhang. Stochastic linear quadratic optimal control with constraint for discrete-time systems. Applied Mathematics and Computation, 2014, 228: 264 – 270.
Xikui LIU received the M.Sc. degree
[26] R. Penrose. A generalized inverse for matrices. Mathematical Proceedings of the Cambridge Philosophical Society, 1955, 51(3): 406 – 413.
Huazhong University of Science and Technology, China, in 2000 and 2004, respec-
[27] A. Albert. Conditions for positive and nonnegative definiteness in terms of pseudo-inverse. SIAM Journal on Applied Mathematics, 1969, 17(2): 434 – 440.
from Shandong University of Science and Technology, and the Ph.D. degree from
tively. He is a associate professor of Shandong University of Science and Technology. His interests include graph theory and DNA computing. E-mail:
[email protected].